Grow. Sunlight grows calories, LEDs grow greens, robots move trays and scout at night; humans still pick fruit.

GROW-S1 · function

Photon supply: daylight first, LED top-up

Deliver the daily light integral (DLI, mol/m²/day) each crop needs at the lowest kWh and capex per mol.

How it works

Photosynthesis is driven by photon count (400–700 nm), so a light recipe is a DLI target = PPFD (µmol/m²/s) x photoperiod (s) / 1e6. Electric lighting energy per m² is DLI / photosynthetic photon efficacy (PPE, µmol/J): at DLI 15 and 3.0 µmol/J a lettuce tier needs about 1.4 kWh/m²/day, or about 7–11 kWh per kg of lettuce at 48–70 kg/m²/yr, before HVAC. Physics caps the gain from better LEDs: the best fixtures reach about 3 µmol/J and the practical ceiling is about 3.4 (white+red) to 4.1 (blue+red) µmol/J, so the remaining lever is yield per photon (canopy capture, far-red, harvest index) and using free sunlight. Greenhouses buy far less electricity per kg than plant factories because the sun supplies the photons; plant factories win on water, CO₂ and land, not on purchased energy.

GROW-S2 · function

Root zone, water and nutrient dosing

Keep oxygen, water and ions at the root at set-points with near-zero discharge.

How it works

Hydroponic NFT channels run a 1–3 mm film of nutrient solution down a slope so roots sit in air and water at once; deep-water rafts buffer temperature and pump failures; aeroponics fogs roots and maximises oxygen but has no buffer if a nozzle clogs. An EC/pH controller doses stock A/B and acid in small pulses, while ion-selective or periodic lab assays correct drift in K, Ca and NO3 that EC cannot see. Transpired water is condensed by the HVAC coil and returned, which is why sealed farms report very high water recirculation. Porous-tube and wick systems survive micro-gravity and power loss but are harder to keep evenly wet: NASA Veggie's passive wicking needed manual watering while APH's porous tubes worked well.

GROW-S3 · function

Crop brain: digital twin and set-point control

Turn sensor data into climate, light, irrigation and harvest decisions without a resident grower.

How it works

A crop growth model (photosynthesis-respiration-partitioning) is coupled to a greenhouse energy balance; model-predictive control (MPC) optimises set-points over a horizon against price and weather forecasts, while reinforcement learning (RL) learns policies in a simulator. The Wageningen Autonomous Greenhouse Challenge is the only public, controlled benchmark: teams remotely set points for real 96 m² compartments, and actuators were driven by a standard process computer. In 2018 one AI team beat expert growers on net profit (EUR 24.78 vs 21.18 per m²) with less heat and water but more electricity; in the 2022 lettuce edition the winner beat the reference, which 'would have been second'. Cameras (RGB/NIR/thermal) give canopy cover and stress indices, so the twin updates its harvest-date estimate daily.

GROW-S4 · function

Plant logistics: bring the plant to the robot

Move trays/towers/gutters through fixed stations (seed, transplant, space, inspect, harvest, wash) so that robots are stationary and cheap.

How it works

The industrial insight in Spread, IGS, AeroFarms and patented tower farms is to turn farming into intralogistics: plants live on standard carriers (trays, towers, gutters) that conveyors, shuttles or AMRs index past stationary stations. A stationary harvester or transplanter works on a known pose with a fixed camera, so it needs no free-space navigation and can be calibrated once. Spacing is done by moving plants from dense to wide-pitch carriers at set ages, keeping canopy light capture high. The cost is a rigid, capital-heavy plant: every carrier, rail and station must be built before the first kg. Iron Ox moved whole grow modules with a mobile robot (Grover) to a robotic workstation; it shut down in 2023.

GROW-S5 · function

Manipulation: harvest, pollinate, de-leaf

Detect, reach, detach and place fruit, and set fruit where bees are impossible.

How it works

A harvest cycle has five steps: detect (RGB-D/CNN), localise the peduncle, plan an obstacle-free approach, detach (cut, twist, or suction-then-cut), and place the fruit. A 2014 review of 50 robots found average success of 66% and 33 s per fruit, with none commercialised; SWEEPER (2020) reached 61% only in a crop trained for the robot and 18% in a normal crop, at 24 s per fruit, half of it logistics. The lesson is that crop architecture (fruit hanging clear of leaves, fixed gutter height, single-truss varieties) matters more than the arm. Flying tethered pickers (Tevel) take the arm to orchard fruit; greenhouse tomato robots (Four Growers) and strawberry robots (Dogtooth, Agrist for peppers) are in early commercial deployment. Strawberries indoors still rely on bees (Oishii).

GROW-S6 · function

Protection and sensing robots (UV-C, laser, scouting)

Replace sprays and scouting walks with light doses and cameras.

How it works

UV-C (about 254 nm) damages fungal DNA; applied at night it defeats photolyase repair that blue light triggers in daylight, so a slow night pass suppresses powdery mildew without residues. Saga Robotics' Thorvald does this on strawberries and vines while also counting flowers and fruit every few days. Laser weeders use a CNN to locate each weed meristem and fire a diode laser for milliseconds, heating it until it dies; throughput scales with the number of laser heads. The same camera pass is a phenotyping scan, so protection and sensing share the platform.

GROW-S7 · function

Open-ground light robots

Seed, weed, water and scout outdoor beds without heavy tractors.

How it works

Cartesian gantries (FarmBot) know each plant's x,y coordinate because they planted it, so watering and weeding are just coordinate lookups. Swarms of small electric seeders (Fendt Xaver: 50 kg, 400 W robots in groups of 6–12) trade width for number and record every seed's position and time, which later becomes the weeding map. Large autonomy retrofits (John Deere, Monarch) keep the tractor form; Monarch's 2025 crisis and lawsuit over tractors 'unable to operate autonomously' show the gap between demo and unsupervised field work.

GROW-S8 · function

On-site protein and loop closure

Turn residues, CO₂ and light into protein; close N, P and carbon loops.

How it works

Black soldier fly larvae (BSFL) digest wet organic residues in 12–14 days and convert roughly 10–26% of substrate dry matter into larval biomass, leaving frass as fertiliser. Duckweeds (Lemna, Wolffia) double fast on shallow nutrient water and are harvested by skimming; growth rises with light up to at least 150 µmol/m²/s. Cyanobacteria (Limnospira/spirulina) in air-lift photobioreactors make protein and O2 from CO₂, as in MELiSSA compartment 4a. These shorten the food chain by one trophic step versus animals, but only BSFL and mushrooms run on waste rather than electricity.